Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.

2, June 2011

ANTERIOR OSTEOPHYTE
IDENTIFICATION IN CERVICAL
VERTEBRAE

Ms. A. T. Chougale1 and Mr. N. B. Sambre. 2

1
Department of Electronics & Telecommunication Engineering, K.I.T.’s college of
Engineering, Kolhapur, Maharashtra.
ashwinishinde2112@gmail.com
2
Department of Electronics & Telecommunication Engineering, K.I.T.’s college of
Engineering, Kolhapur, Maharashtra.
sambrenitin@yahoo.com

ABSTRACT
Radiologist always examines X-ray to determine abnormal changes in cervical, lumbar & thoracic
vertebrae. Osteophyte (bony growth) may appear at the corners of vertebrae so that vertebral shape
becomes abnormal. This paper presents the idea from Image processing techniques such as customised
Hough transform which will be used for segmentation which should be independent of rotation, scale,
noise & shape. This segmented image will be then used for computing size invariant, convex hull based
features to differentiate normal cervical vertebrae from cervical vertebrae containing anterior
osteophyte. This approach effectively finds anterior osteophytes in cervical vertebrae.

KEYWORDS
Osteophytes, customised Hough transform, convex hull, segmentation.

1. INTRODUCTION
Radiographs of the spine provide a practical approach for detecting and assessing vertebral
abnormalities that may be related to osteoarthritis or osteoporosis. The presence of bony
growths (‘Osteophyte’) on vertebra corners, disc space narrowing are all features commonly
evaluated visually from radiographs.

The presence of osteophytes is significant because it may be related to degeneration in the

attachment of the outer annular fibers of the disc to the vertebral endplate. This degeneration
may allow the vertebra to slip to the anterior or to both the anterior and the side [12]. Vertebra
distortion based on anterior osteophytes results in a deviation from the characteristic rectangular
vertebral shape, which tends to be convex. This proposed work focuses on vertebral distortion
along the anterior boundary as an indicator of Osteophyte. Fig. 1 presents three examples of
anterior osteophytes in vertebrae with osteophytes in the upper left-hand corner (a), osteophytes
in the lower left-hand corner (b) and osteophytes in both left-hand corners (c). Fig. 3 shows a
normal vertebra (left-hand side is the anterior side).

DOI : 10.5121/sipij.2011.2207 71
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

Figure.1. Three examples of anterior osteophytes vertebrae with osteophytes The left-hand side
is the anterior side of the vertebra shown in upper left-hand corner (a), osteophytes in the lower
left-hand corner (b),and osteophytes in both left-hand corners (c)

Fig.2. Example for normal vertebra without anterior osteophytes

GHT uses an edge image to correlate points in a previously defined template to those in the
target image using local gradient information. If a good edge image is not available, too few
votes will be correlated to the template [1]. GHT assume that a clear edge image, from which
gradient information can be obtained. Since x-ray images are not clear edge image, an edge
detection process is used. The proposed process performs Gaussian blurring on a copy of the
original image and the result is subtracted from the original. From the resulting edge image,
extraction of the gradient information will be performed using optimum gradient operators.

In prior research, image processing techniques were investigated to compute features along the
anterior boundary of cervical spine vertebrae for differentiating normal vertebrae from vertebrae
with abnormal osteophytes [13]. The features examined were radius-of-curvature-and grayscale
gradient-based features computed along the anterior boundary of cervical spine vertebrae. The
radius of curvature-based features highlights the relative constriction along the anterior
boundary between normal and abnormal vertebrae. The grayscale border gradient features
examine the grayscale contrast difference between the vertebral interior and exterior along the
boundary. The primary difficulty with the radius-of-curvature features was that they were not
adaptive to vertebral size. In this research, size-invariant convex hull features are introduced to
detect anterior osteophytes in cervical spine vertebra. Size-invariant features are investigated in
order to facilitate the direct comparison of vertebrae between radiographs.

In this proposed work, size-invariant convex hull features will be used to detect anterior
osteophytes in cervical spine vertebra. Size-invariant features will investigate for the direct
comparison of vertebrae between radiographs. The convex hull of a data set of points is the
72
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

smallest convex set of points that includes the data set of points. The convex hull will
determined from a vertebra’s boundary points in order to compute features characterizing the
vertebra’s shape. Examining Figs.1 and 2, a normal vertebra generally has rectangular shape and
tends to be convex. Thus, the convex hull for a cervical spine without anterior osteophytes is
anticipated to be similar to its original shape on the anterior side. Inspecting Fig. 1, the
protrusion region along the anterior side of the vertebra contains osteophytes, making the
vertebra shape concave. Four size-invariant features are determined based on comparing a filled
vertebra with its filled convex hull.

A standard multi-layer perceptron (MLP) will be used to evaluate the features for osteophytes
discrimination on a substantial lumbar vertebrae data set [2].

2. ALGORITHM FOR OSTEOPHYTES IDENTIFICATION IN

CERVICAL VERTEBRAE

In this proposed work, image analysis techniques are investigated for detecting abnormal (i.e.
having osteophytes) cervical spine vertebrae. For individual vertebra analysis automatic
vertebral segmentation will be performed. The steps for automatic vertebra segmentation,
vertebral feature calculations, and vertebral discrimination are presented in following sections.

2.1. VERTEBRAL BOUNDARY DETERMINATION:

The customized Hough Transform will be used to captures shape variability and exploits shape
information embedded in the Hough domain to overcome noise. The following subsections
describe the different processing steps involved in the CHT [10].
2.1.1. Extraction of gradient information:

Here unsharp masking is proposed to obtain edge image. Mathematically the resulting image
after unsharp masking can be represented as:
Fu(x, y) = f(x, y) - fb(x, y) (1)
Where f(x, y) represents the original image [Figure.3 (a)], fb(x, y) represents the Gaussian
blurred image [Figure.3 (b)], and fu(x,y) represents the unsharp masked image [Figure.3(c)].
Here subtraction of a blurred image from the original will provide only abrupt local variations in
edges and eliminate an unnecessary information (background). An averaging filter is used to
reduce the unwanted high frequency components. Then the resulting edge image is binaries.
Finally gradient operators are passed to obtain edge image. The Sobel and Prewitt operators can
be used to estimate the gradient information but there is inconsistency between the discrete and
continuous gradient information. This can be minimized using the gradient descent method [6].
The optimum 3 x 3 gradient operator is given as:

0.112737 0.274526 0.112737

∇x = ∂
∂x = 0 0 0 

0.112737 0.274526 0.112737
0.112737 0 0.112737
∇y = ∂
∂y = 0.274526 0 0.274526 (2)
0.112737 0 0.112737

73
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

(a) (b) (c)

Figure 3. Unsharp masking (a) Original image (b) Gaussian smoothed image and (c) image after
unsharp masking

It was found that the 3 x 3 optimum gradient operators yield the highest peak in the accumulator
and the most compact and uniform region surrounding a possible candidate for the reference
point.

2.1.2 Definition of template:

GHT is a template-based technique. It looks for instances of previously defined template in the
target image. Even if good gradient information is available, the template must adequately
represent the target object in order to obtain the necessary number of votes in the accumulator.
The use of multiple templates should capture shape variability and closely represent the
different classes of shapes seen across the data set. However, the use of a collection of templates
may be computationally expensive. Consequently, the selection of a single template that
represents the target object is of primary interest. As a first approximation, we have used the
mean of 50 templates that were obtained from the manual land marking of 50 images [6]
provided by the second National Health and Nutrition Examination survey (NHANES II)
database [11]. The mathematical representation of the templates and that of the mean template is
as follows:

(Template)i = [ xi1, …., xi80, yi1, …., yi80 ]^T

1
∑
i = 50
mean_template = (Template)i (3)
50 i =0

2.1.3. Accumulator updating and reckoning of votes:

Updating the accumulator, as well as picking the best estimate for the reference point is the
remaining part of the algorithm.

A. Hough Domain update process:

Here, template matching is done based on the gradient information and for each match the
accumulator bin is updated and this search and update process is repeated for Every value of
scale‘s’& rotation ‘Φ’.A look-up table (R-table) for the template edge image is created with
different values of (x, y), where r is the distance from the edge to the chosen reference point and
α is the angle between the radius vector and the horizontal. The R-table summarizes the shape
information of the template and is a part of the parametric representation of the image. Now,

74
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

with the R-table, template matching is done on the target image and votes for possible matches
are collected in the accumulator structure.

B. Reckoning of votes in Hough Domain:

The accumulator is a 4-D structure where 2 of the dimensions (m x n) represent the spatial
coordinates and the remaining two (j x k) represent scale and rotation variations of the template.
As a consequence, the 4-D accumulator can be seen as a (j x k) collection of 2-D images of the
spatial location of the reference and each of those images can be treated individually. If the 2-D
accumulator is plotted for a fixed value of scale and rotation [Fig. 2(a)], we would expect as per
theory, a well defined peak. In practice, the peak is neither unique nor well defined. Votes do
not form uniform and compact regions and this makes the task of finding a good candidate for
the reference point very difficult. Excessive noise in the original images makes votes sparse in
the accumulator, yielding non-uniform structures. To fix this problem, Gaussian filtering is
performed and this greatly improves the chances of finding an absolute maximum peak and this
leads to a more relevant observation: Several peaks tend to align describing the orientation and
location of the cervical vertebrae. The neighborhood peaks represent a more complete figure of
merit about the overall performance of the template. A small window, having the highest peak
as the central point is defined as our region of interest. The next step is to try and find the
approximate orientation of the spine and the location of the neighboring peaks with the
knowledge of the location of the highest peak. Let γ be the angle that approximates the
orientation of the template and ‘Φ’ the current rotation angle of template. Then ‘µ’defined as

µ=γ+Φ (4)

gives a good approximation of the orientation of the cervical spine in the target image. This
angle is then used to rotate the small window as shown in Fig. 2(b). An intensity profile passing
through the center of the window is defined as the characteristic profile and this has the property
of passing through the center of the highest peak and if the rotation angle is correct, it should
also pass through the neighboring peaks. A plot of the characteristic profile can be seen in Fig.
2(c). Once the highest peak has been found, the votes corresponding to this peak are eliminated,
so that the next search yields the second highest peak. The same process is repeated and the
third peak is also found. The output of the process will be Â, the cumulative value of the highest
peak and its 2 neighbors. This step is a quantitative measure of how well the template matches
the whole cervical vertebrae. The next subtlety involved in this analysis is the shape of the
peaks. A blurred peak will indicate that in spite of the considerable quantity of votes that form
the peak; the template matching is poor, yielding a widespread structure. On the other hand, a
sharp peak will indicate a more accurate match since most of the votes concentrate in that small
region. Hence, we use the gradient operator to measure the decay rate of the highest peak, and
refer to this measure as Ĝ
The new criterion for reckoning of votes can be defined as a weighted sum of two quantities

Proposed criteria = C (Â) +ε (Ĝ) (5)

where β & ε are constant less than one.

For every value (Φ, s), a 2D image is created and for each of those we compute a pair of
corresponding (Â, Ĝ) values and store them. The (x, y) location of the highest peak is stored as
the best candidate for the reference point. While the traditional criteria chooses the best (Φ,s)
pair to be the highest value in the accumulator, the new criterion for reckoning of votes chooses
the best (Φ,s) pair using both  ,Ĝ values as follows:

75
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

1. Find the highest value in the accumulator for each value of scale and orientation.

2. Define the subset of each best (Φ,s)pair:

(Φ,s)best={(Φ,s)Â(Φ,s)≥β.Â(Φ,s)}

2.2. VERTEBRAL FEATURES CALCULATION:

In this section, four convex hull-based features are investigated for vertebra identification [2].
The following notation is used for defining the features. Let I denote the M x N intensity
cervical spine X-ray image, where1 ≤ x ≤M and 1 ≤y ≤ N. Let V denote a cervical vertebra
within an X-ray image with area Av such that

V= 1 if(x; y) is on or inside the vertebra boundary

= 0 otherwise

The convex hull for V is determined using the quick convex hull algorithm [9]. Let C denote the
resulting filled convex hull for vertebra V such that

C = 1 if(x; y) is on or inside the convex hull boundary for

= 0 otherwise

Let E denote the set of exclusive-OR points between V and C such that

E(x, y) =V(x, y) ⊕ C(x, y). E is expected to contain one or more connected regions, for each
concave vertebral side. Using 8-connectivity, let k denote the number of unique connected
components within E. The four features used for detecting anterior osteophytes include: (1) the
ratio of the vertebral area to the convex hull area, (2) the ratio of the exclusive-OR area to the
convex hull area, (3) the ratio of the exclusive-OR area on the vertebra’s anterior side to the
vertebra area and (4) the ratio of the area of the largest connected region from the exclusive-OR
regions on the anterior side of the vertebra to the vertebra area.

The goal with feature extraction is to highlight protrusion regions on the anterior side of the
vertebra that are characteristic of osteophytes. The ratios of the exclusive- OR area to the filled
convex hull and the vertebra area, respectively, are intended to quantify the variation of the
vertebral shape from the vertebra’s ideal rectangular shape. The exclusive-OR regions represent
concave regions around the vertebra’s periphery. The final exclusive-OR-based feature focuses
on the anterior side of the vertebra for finding the largest concave region and comparing the size
of the concavity to the vertebra size in order to highlight the location of the largest protrusion
region. The features examined provide size-invariant measures of the deviation of the filled
convex hull from the original filled vertebra. The purpose for exploring size-invariant features is
to enable the direct comparison of different vertebrae within the same X-ray image (i.e. the
same patient) and vertebrae from different images (i.e. different patients).

76
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

Figure (4): The exclusive-OR operation between original vertebra and its convex hull image for
a normal vertebra.

Figure (5). The exclusive-OR operation between original vertebra and its convex hull image for
an abnormal vertebra with anterior osteophytes(lower left-hand side).

For the initial feature, the ratio of the vertebral area to the filled convex hull is denoted as R and
is defined as R = AV/AC: The ratio of the exclusive-OR area to the filled convex hull area is
denoted as T and is defined as T =AE/AC: In order to illustrate this feature, Figure(4) presents
the exclusive-OR for a normal lumbar vertebra, and Figure(5) provides the exclusive-OR of an
abnormal vertebra with anterior osteophytes on the lower left-hand side. From Figure (4) and
(5), there are a number of disconnected regions in the exclusive-OR between the filled vertebra
and its corresponding filled convex hull.

The following procedure is used to compute the final two exclusive-OR-based features. The
centroid location, (Xc, Yc), is calculated for the vertebra V. Second, the connected regions
within E are connected component labeled. Third, the original XY plane is shifted to new X1Y1
plane for V, which is centered at the centroid location with the axes parallel to the original XY
axes.

77
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

Figure(6). The exclusive-OR area after removing the posterior side of a normal vertebra.
(Example for a normal vertebra).

Figure (7). The exclusive-OR area after removing the posterior side of an abnormal vertebra

The third feature computed is the ratio of the exclusive-OR area on the vertebra’s anterior side
to the vertebral area. Connected components in E that touch or are entirely located on the right-
hand side of the X1Y1 plane (i.e. positive X1 half-plane) are removed. In other words, any
connected component in E with one or more pixels on the Y1 axis or the negative X1 axis are
removed. These connected components are considered as part of the posterior portion of the
vertebra and are not related to anterior osteophytes. Let G denote the number of connected
components remaining in the negative X1 plane (i.e. anterior side). Let H = {h1 h2… hG}
refer to the set of connected components remaining with areas given by

Ah; for 1 ≤ i ≤ G: Then, the ratio of the exclusive-OR area on the vertebra’s anterior side to
G
vertebral area is denoted as U and is defined as U = ∑i=1
Ahi / Av. The final convex hull-based

feature is based on finding the largest connected component on the anterior side. Then, the ratio
of the largest connected region from the exclusive-OR regions on the anterior side of the
vertebra to the vertebral area is denoted as F and is defined as F = maxi (Ahoi)/AV;
where 1 ≤ i ≤ G: Note that G ¼ 1 for the lumbar vertebrae examined. Figures6 and 7 show
examples of the exclusive-OR area after removing the posterior portion of the image for normal
and abnormal vertebrae, respectively.

78
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

2.3. RELATED WORK:

Abraham Tezmol, Hamed sari-sarraf , Sunanda Mitra, Rodney Long, Arunkumar
gururajan perform customized hough transform for robust segmentation of cervical
vertebrae from x-ray images.The aim of customized Hough transform (CHT), that
captures shape variability and exploits shape information embedded in the accumulator
structure to overcome noise and occlusions is proposed. This approach effectively finds
estimates of the location and orientation of the cervical vertebrae boundaries in digitized
x-ray images.

Anterior osteophytes discrimination in lumbar vertebrae using size-invariant features,

Image processing techniques are presented for computing size-invariant, convex hull-
based features to highlight anterior osteophytes by Maruthi Cherukuria, R. Joe Stanleya,
Rodney Longb, Sameer Antanib, George Thomab

In this proposed work Customized Hough transform will be used for segmentation &
size-invariant, convex hull-based features to locate anterior osteophytes in cervical
vertebrae

3. CONCLUSIONS
In this research, image analysis techniques are presented for anterior osteophytes identification
in cervical vertebrae. The task of segmenting the cervical vertebrae has been carried out using a
customized Hough transform. This technique deals with shape information present in the voting
structure. The gradient information & template representation are very important aspects. For
better shape variability more templates are used, approximately 50 templates give 100% shape
variability. Four size-invariant, convex hull based features will be analyzed for differentiating
normal vertebrae from abnormal vertebrae (containing anterior osteophytes). The convex hull-
based features provide the variation in vertebra’s shape from a typical convex shape, showing
the variation on the anterior side of the vertebra. The ratio of the largest connected component
on the anterior side from the exclusive-OR between the filled vertebra and its filled convex hull
denoted as F. This will provide the highest and most consistent results of the four convex hull-
based features.
Furthermore, the features investigated are size-invariant and can be applied to vertebra analysis
in different X-ray images.

REFERENCES
[1] D.H. Ballard, “Generalizing the Hough Transform to detect arbitrary shapes”, Pattern Recognition,
13(2), pp. 111-122, 1981.

[2] Maruthi Cherukuri “Anterior osteophyte discrimination in lumbar vertebrae using size-invariant
features” Computerized Medical Imaging and Graphics 28 (2004) 99–108

[3] V.F. Leavers, “Which Hough Transform?”, Image Understanding, Vol. 58, No.2, pp. 250-264,
Sep.1993.

[4] H. Li, M. A. Lavin, and R. L. LeMaster, “Fast Hough Transform”, Proceedings 3rd Workshop on ]
Computer Vision, Bellair, MI, 1985.

79
 Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

[5] J. Illingworth and J. Kittler, “The adaptive Hough Transform”, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 9 (5), 1987.

[6] S. Ando, “Consistent gradient operators”, IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 22, No.3,

[7] Taylor JA, Resnick D. The aging spine: radiographic-pathologic correlation. In: Genant HK, Jergas
M, van Juijk C, editors. Vertebral fracture in osteoporsis, research group. San Francisco, CA:
University of California at San Francisco; 1995.

[8] Stanley RJ, Long R. A radius of curvature-based approach to cervical spine vertebra image analysis.
Copper Mountain, CO. Proc 38th Annu Rocky Mountain Bioengng Symp 2001;37:385–90.

[9] Barber CB, Dobkin DP, Huhdanpaa HT. The Quickhull algorithm for convex hulls. ACM Trans Math
Software 1996;22(4):469–83.Indian Institute of Science, Bangalore,India, Jan. 1999.

[10] AbrahamTezmol “Customized Hough Transform for Robust Segmentation of Cervical

Vertebraefrom X-Ray Images” Fifth IEEE Southwest Symposium on Image Analysis and
Interpretation (SSIAI.02) 0-7695-1537-1/02$17.00©2002IEE

[11] Long LR, Thoma GR. Image query and indexing for digital X-rays.SPIE Conference on Storage and
Retrieval for Image and Video Databases VII, San Jose, CA. Proc SPIE 1999;3656:12–21.

[12] TAYLOR JA, RESNICK D. THE AGING SPINE: RADIOGRAPHIC-PATHOLOGIC CORRELATION. IN: GENANT
HK , JERGAS M, VAN JUIJK C, EDITORS. VERTEBRAL FRACTURE IN OSTEOPORSIS, RESEARCH GROUP.
SAN FRANCISCO, CA: UNIVERSITY OF CALIFORNIA AT SAN FRANCISCO; 1995.

[13] Stanley RJ, Long R. A radius of curvature-based approach to cervical spine vertebra image analysis.
Copper Mountain, CO. Proc 38th AnnuRocky Mountain Bioengng Symp 2001;37:385–90.

[14] Genant HK, Ettinger B, Harris ST, Block JE, Steiger P. Quantitative computed tomography in
assessment of osteoporosis. In: Riggs BL,Melton LJ, editors. Osteoporosis: etiology, diagnosis and
management.New York: Lippincott Williams & Wilkins Publishers.1988:221–49.

[15] Stanley RJ, Long R. A radius of curvature-based approach to cervical spine vertebra image analysis.
Copper Mountain, CO. Proc 38th Annu Rocky Mountain Bioengng Symp 2001;37:385–90.

[16] Barber CB, Dobkin DP, Huhdanpaa HT. The Quickhull algorithm for convex hulls. ACM Trans
Math Software 1996;22(4):469–83.

80